PaperlayerView exact revision on arXiv

Source-linked AI summary

Doc2EDAG: An End-to-End Document-level Framework for Chinese Financial Event Extraction

Shun Zheng, Wei Cao, Wei Xu, Jiang Bian

arXiv:1904.07535v2cs.CLcs.LG

TL;DR

Sentence-level event extraction is limited when document arguments scatter across sentences and multiple event records coexist. Doc2EDAG directly generates an entity-based directed acyclic graph for document-level extraction, using no-trigger-word labeling and a large Chinese financial-announcement dataset. Experiments report superiority over state-of-the-art methods on these document-level challenges.

  • Problem

    Sentence-level event extraction overlooks arguments scattered across document sentences and struggles when multiple event records coexist, motivating document-level extraction.

  • Method

    Doc2EDAG transforms event tables into entity-based directed acyclic graphs, uses document-level contexts and memory for path expansion, and reformalizes DEE without trigger words.

  • Results

    Doc2EDAG significantly outperforms state-of-the-art methods on DEE-specific challenges in experiments on a large real-world Chinese financial-announcement dataset.

  • Takeaways & Limitations

    The paper's general labeling and modeling strategies are presented as applicable beyond finance to business domains with similar document-level challenges.

  • Takeaways & Limitations

    The evaluation focuses on pure text sequences; extending Doc2EDAG to richly formatted inputs remains future work.

Abstract

from arXiv · show

Most existing event extraction (EE) methods merely extract event arguments within the sentence scope. However, such sentence-level EE methods struggle to handle soaring amounts of documents from emerging applications, such as finance, legislation, health, etc., where event arguments always scatter across different sentences, and even multiple such event mentions frequently co-exist in the same document. To address these challenges, we propose a novel end-to-end model, Doc2EDAG, which can generate an entity-based directed acyclic graph to fulfill the document-level EE (DEE) effectively. Moreover, we reformalize a DEE task with the no-trigger-words design to ease the document-level event labeling. To demonstrate the effectiveness of Doc2EDAG, we build a large-scale real-world dataset consisting of Chinese financial announcements with the challenges mentioned above. Extensive experiments with comprehensive analyses illustrate the superiority of Doc2EDAG over state-of-the-art methods. Data and codes can be found at https://github.com/dolphin-zs/Doc2EDAG.

1 Introduction

Financial and other document collections create document-level event-extraction challenges because event arguments can scatter across sentences and multiple event records can coexist. Doc2EDAG addresses these challenges with end-to-end EDAG generation and no-trigger-word document-level labeling, evaluated on a larger, more multi-event financial-announcement dataset.

  • Motivation: Document-level event extraction matters for growing financial, legal, and health document collections that contain valuable structured information.The paper motivates EE for extracting information from large digital document collections.
  • Challenges: Arguments-scattering means one event's arguments may appear across multiple sentences, while multi-event documents contain several event records.The Equity Pledge example illustrates both challenges and shows that sentence-level choices can be wrong at the document level.
  • Prior limitations: Most prior sentence-level EE work uses ACE 2005-style annotations limited to arguments within a sentence, leaving arguments-scattering insufficiently addressed.The paper distinguishes this sentence-level setting, SEE, from document-level EE required for financial announcements.
  • Prior limitations: DCFEE combines sentence-level sequence tagging, key-event-sentence detection, and heuristic argument padding, but sequence tagging handles multi-event sentences poorly and context-agnostic completion misses scattered arguments.These limitations motivate an end-to-end document-level approach.
  • Proposed approach: Doc2EDAG transforms event tables into entity-based directed acyclic graphs, decomposing table filling into sequential path-expansion subtasks with document-level entity contexts and memory.The model directly generates event tables from documents and avoids trigger-word labeling.
  • Evaluation: The paper reformalizes DEE without trigger words and builds a real-world financial-announcement dataset that is ten times larger than DCFEE's data, with about 30% of documents containing multiple event records.The dataset and experiments target arguments-scattering and multi-event challenges.
  • Evaluation: Extensive experiments show that Doc2EDAG significantly outperforms state-of-the-art methods on DEE-specific challenges.The introduction reports this as the central empirical conclusion.

2 Related Work

Prior EE work largely targets sentence-level extraction, while document-level settings require handling scattered arguments and multiple event records. Doc2EDAG builds on joint extraction and distant-supervision research to address these document-level challenges.

  • Joint extraction research simultaneously models entities and structures such as relations or events.
  • Sentence-level EE methods overlook arguments that are scattered across multiple sentences.
  • Distant supervision reduces reliance on expert annotations, but conventional EE requires trigger-word labeling resources.
  • Combining distant supervision with simple constraints can produce good document-level event-labeling quality because knowledge bases and documents share a domain and events contain multiple arguments.

3 Preliminaries

The paper defines event-extraction units and uses ChFinAnn, a large collection of official Chinese stock-market disclosures, while focusing on frequent and influential event-related documents.

  • An event role is a predefined field of an event table.
  • An event argument is an entity that plays a specific event role, and an event record contains arguments for required roles.
  • ChFinAnn contains official disclosures from listed Chinese companies, spanning hundreds of document types.
  • The study focuses on event-related document types that are frequent, influential, and mainly expressed in natural language.

4 Document-level Event Labeling

The paper labels document-level events by matching event-knowledge-base records to documents under constraints, then formulates DEE as trigger-free event-table filling.

  • Document-level event labeling maps tabular knowledge-base records to document text and treats well-matched records as expressed events.
  • Matched records must contain predefined key-role arguments and exceed an event-specific argument-count threshold.
  • The no-trigger-words design labels arguments and roles without explicitly labeling trigger words.
  • Avoiding trigger-word labels simplifies distant-supervision labeling by removing dependence on trigger dictionaries or manually curated filtering heuristics.
  • Trigger-free DEE comprises entity extraction, event detection for each event type, and event table filling.

5 Doc2EDAG

Doc2EDAG encodes document-level entities and generates entity-based directed acyclic graphs through sequential role-wise path expansion. Its pipeline combines entity recognition, event triggering, memory-based expansion, and joint optimization.

  • Entity Recognition: Entity recognition uses a Transformer encoder with a CRF layer and BIO labels derived from matched argument roles.
  • Document-level Entity Encoding: Document-level encoding exchanges information among entity mentions and sentences, adds sentence-position embeddings, and merges mentions sharing a surface name.
  • EDAG Generation: For each triggered event type, path expansion predicts whether each entity should be added using path history, entity embeddings, and a role-indicator embedding.
  • EDAG Construction: Doc2EDAG transforms event records into an entity-based directed acyclic graph whose complete paths correspond to event-table rows.
  • EDAG Generation: The model decomposes table filling into sequential path-expanding subtasks ordered by predefined event roles.
  • Optimization: The final objective sums entity-recognition, event-triggering, and EDAG-generation losses with tunable weights.
  • Inference: At inference, the model recognizes entities, encodes document context, and generates an EDAG through sequential path expansion.
  • Practical Tips: Scheduled sampling reduces the training–inference discrepancy, while weighting negatives more heavily addresses harmful false-positive path expansions.

6 Experiments

Experiments evaluate Doc2EDAG on a large Chinese financial-announcement dataset designed for document-level event extraction, including arguments-scattering and multi-event cases. Doc2EDAG substantially outperforms baselines, with document-level modeling and its major components contributing to performance.

  • 6.1 Experimental Setup: The dataset contains 32,040 documents, is ten times larger than DCFEE’s, and is split chronologically into train, development, and test sets at 8:1:1.The dataset covers 2008–2018 Chinese financial announcements and five event types.
  • 6.1 Experimental Setup: The evaluation directly compares predicted and ground-truth event tables using micro-averaged role-level precision, recall, and F1 as the event-level metric.Statistics are aggregated across documents after matching predicted and ground-truth records without replacement.
  • 6.2 Performance Comparisons: Doc2EDAG improves over DCFEE-O by 19.1, 4.2, 26.5, 28.4 and 13.4 F1 scores on EF, ER, EU, EO and EP events, respectively.It achieves significant improvements over all baselines for all event types.
  • 6.2 Performance Comparisons: On the multi-event set, Doc2EDAG increases average F1 by 17.7 points over DCFEE-O, while all models’ performance drops significantly.Doc2EDAG maintains the highest extraction performance across the evaluated scenarios.
  • 6.3 Analysis: Removing the path memory causes the largest declines, scheduled sampling improves average F1 by 5 points, and document encoding contributes 2.1 average F1 points.A larger negative class weight provides slight but stable gains across event types.
  • 6.4 Case Studies: Doc2EDAG recovers the correct EDAG in the case study, whereas DCFEE makes many mistakes even with a perfect sentence-level extraction model.The case study illustrates the benefit of end-to-end document-level modeling.

7 Conclusion and Future Work

The paper concludes that Doc2EDAG combines end-to-end document-level event extraction with no-trigger-word task formalization and demonstrates effectiveness on a large financial dataset. Future work will extend its inputs beyond pure text sequences to richly formatted documents.

  • 7 Conclusion: Doc2EDAG provides end-to-end document-level event extraction together with a no-trigger-word DEE formalization for distant-supervision labeling.The paper also builds a large-scale real-world financial dataset and conducts extensive empirical studies.
  • 7 Conclusion: The general labeling and modeling strategies use no domain-specific assumption and can benefit practitioners in other domains directly.The stated examples include legal and medical document extraction.
  • 7 Future Work: Future work will explore expanding Doc2EDAG inputs from pure text sequences to richly formatted ones.This extension is presented as appealing but left for future investigation.

A Appendix

The appendix supplies omitted methodological, implementation, and evaluation details, including event specifications, hyperparameters, EDAG pseudocode, and additional studies.

  • A.1 Event Type Specifications: Appendix Section A.1 presents event type specifications and corresponding preprocessing details.
  • A.2 Hyper-parameter Setting: Appendix Section A.2 reports the hyperparameter settings used to run the experiments.
  • A.3 EDAG Generation: Appendix Section A.3 provides pseudocode to facilitate understanding of EDAG generation.
  • A.4 Additional Evaluation: Appendix Section A.4 adds evaluation results for entity extraction and event triggering.These are preceding subtasks before event-table filling.
  • A.5 Case Studies: Appendix Section A.5 studies three further sophisticated cases illustrating the necessity and advantages of end-to-end modeling.

A.1 Event Type Specifications

Event labeling marks key roles that must be non-empty and imposes event-type-specific minimum numbers of matched roles. These constraints are empirical and adjustable for other domains to trade precision against recall.

  • Event Type Specifications: Key roles are required to be non-empty during document-level event labeling.
  • Event Type Specifications: The minimum numbers of matched roles for EF, ER, EU, EO and EP are 5, 4, 4, 4 and 5, respectively.
  • Event Type Specifications: Practitioners in other domains can adjust these empirical constraints to trade off precision and recall according to task requirements.

A.2 Hyper-parameters Setting

The paper reports the hyper-parameter settings for reproducibility and orders event roles by their non-empty argument ratios during training.

  • Hyper-parameters are summarized in Table 7 to support reproducibility.
  • Training follows decreasing non-empty argument ratios across roles to provide more informative path histories for recurrent decoding.The authors validate this intuition against models trained with randomly permuted role orders.

A.3 Pseudo Codes for the EDAG Generation

The appendix provides EDAG loss and inference pseudocode alongside implementation details and case studies. These materials explain how Doc2EDAG represents event structures and why it addresses scattered arguments and multiple event mentions.

  • Additional Evaluation Results: Table 8 complements end-to-end evaluation with entity-extraction and event-triggering results for each event type.
  • Case Studies: The case studies illustrate errors involving multiple events, duplicated mentions, and arguments scattered across sentences, motivating Doc2EDAG’s end-to-end design.The examples cover Equity Overweight, Equity Underweight, and Equity Freeze events, with incorrect predicted arguments highlighted.
  • Pseudo Codes for the EDAG Generation: The EDAG loss pseudocode initializes loss, iterates over event and role types, and accumulates path-expanding classification losses.
  • Pseudo Codes for the EDAG Generation: During loss computation, each entity node supplies path-expanding labels for classification, while known expanded entities update corresponding role memories.
  • Pseudo Codes for the EDAG Generation: Inference initializes a virtual starting node, processes roles in predefined order, and expands each current leaf with predicted entities.
Loading 1904.07535v2…